Can AI Break the Mythical Man

Editorial Comment

After reading this “future discussion record” from 2026, the most intuitive feeling is that we are finally starting to face the “demolition-style” reconstruction of software engineering’s underlying logic by AI. The rules Fred Brooks established in The Mythical Man-Month have dominated programmers’ perceptions for half a century. The core logic consists of just two points: first, communication costs explode exponentially with the number of people; second, tools can only solve superficial “accidental complexity” and cannot touch the “essential complexity” of requirements. However, this discussion record accurately points out a fact—AI is flanking Brooks’s defenses from both dimensions simultaneously. I strongly agree with a core judgment in the text: AI is not just another better IDE or compiler; it is the first technology capable of touching “essential complexity.” Previously, we said that when a requirement was vague, we had to rely on senior architects to repeatedly align, guess the product manager’s intentions, and avoid edge-case pitfalls; these tasks were considered the “last bastion of human intelligence.” But the current trend is that AI Agents are structuring these The line distinguishing accident from essence begins to blur. The return of the “surgical team” model is, in my opinion, the most brilliant insight of the entire piece. Brooks originally envisioned a superstar leading a group of assistants, which is almost contrary to human nature—who would want to study hard for over a decade just to write documentation and run tests for someone else’s code? The collapse of this organizational structure stems from human self-esteem and professional aspirations. But AI Agents perfectly fill this gap; they have no pressure for promotion, don’t need 1:1 check-ins, and are the most perfect “copilots.” This “1 Architect + N Agents” structure might completely end the bloated product and research teams of hundreds often seen in big tech, reducing communication costs from a headache-inducing socio-psychological problem to an engineering problem that can be solved by increasing computing power and optimizing context. As for the credibility of this record No single technology can increase software productivity tenfold.

His reasoning is: there are two types of difficulties in software development. Accidental complexity (clumsy languages, primitive tools, slow compilation, difficult debugging) and essential complexity (ambiguity of requirements, conceptual complexity of the system, difficulty in mapping specifications to reality). The progress of the past few decades—high-level languages, IDEs, version control, CI/CD—has primarily eliminated accidental complexity. But essential complexity is an inherent property of the problem itself and will not disappear just because tools improve. Therefore, no silver bullet exists.

AI is the first technology to truly make an impact at the level of essential complexity.

All previous tools were helping you “write code faster.” AI is different—it can help you understand unfamiliar domain logic, explore solution spaces you hadn’t considered, provide structured suggestions for vague requirements, and infer edge cases you might have missed when describing an intent. These are not accidental complexity; they are part of essential complexity.

This doesn’t mean a silver bullet has appeared. But it means that the premise of Brooks’ “No Silver Bullet” argument—that tools can only handle accidental complexity—has been partially shaken for the first time in 2026. If AI can continue in ess

ential complexity, Brooks’ "impossible 10x" would need to be re-examined.

Brooks’ surgical team model hasn’t been overturned by AI—instead, AI has finally made it feasible.

In his book, Brooks proposed a solution: The problem he raised (communication costs exploding with headcount) and the solution he proposed (the surgical team) are both correct; it’s just that the solution waited 50 years for a feasible implementation.

The final consensus of previous discussions was that “judgment is the non-parallelizable bottleneck.” But this might be a temporary The real suspense of Brooks 2026 might not be whether "judgment can be parallelized," but whether the concept of "judgment" itself will be gradually decomposed into smaller and smaller sub-capabilities, most of which are eventually automated, leaving only a shrinking irreducible core.

One last angle not previously touched upon: Brooks’s n² communication cost assumes isomorphic participants. AI is not isomorphic.

Brooks says that n people require n(n-1)/2 communication channels because everyone might need to coordinate with everyone else. But the premise of this model is that every participant is an independent agent with autonomous judgment.

The reason communication costs between humans are n² is largely because humans have independent mental states, different understandings, different priorities, and different egos. When you remove these, the mathematics of communication costs changes. It’s not that n² is optimized to n log n, but that the root cause leading to n² no longer exists.

This doesn’t mean AI teams have no coordination costs—as discussed before, token costs, context bloat, and output conflicts are real. But these are engineering problems, not social problems. Engineering problems can be systematically optimized; social problems can only be managed.